Abstract

Confidence measures for the output of a speech recognizer have been, for some years now, a topic of interest in the speech community. Initially the main goal was to use them as diagnostic tools to understand recognizer behavior by identifying regions and sources of error. But they have also proven useful in other tasks, such as supervised or unsupervised acoustic model adaptation, confidence conditioned language modeling, semantic analysis, and improving word accuracy through rescoring techniques [3]. In all these cases various features were used to compute a confidence measure for each hypothesized word or sentence [2, 5]. In this paper we use similar features that can be measured for each hypothesized word-token and have been commonly used in the literature to assign confidence. However we use them to improve recognition performance. We treat them as knowledge sources, and combine them linearly, along with the acoustic and language model log-probabilities, to get a new log-likelihoo...

Citations